Research On Vehicle Path Planning Based On Visual SLAM

For a long time,it's people's demand for extreme environment exploration that has promoted the development of mobile robot technology.These environments are often unknown and full of danger,which requires mobile robots to be able to perceive and recognize the surrounding environment,and achieve autonomous learning and path planning.Path planning of mobile robots refers to a collision-free and end-point path for robots from the starting point.Aiming at unknown environment,this paper uses visual SLAM technology.By using Kinect2 to extract environment data,the construction of environmental map based on RGB-D SLAM is realized.For the problem of path planning,most of the traditional path planning algorithms are based on the known environment model,and are vulnerable to the influence of environmental factors,resulting in large deviations in the actual operation results,so this paper abandons the traditional path planning methods and adopts the path planning algorithm based on reinforcement learning.In this paper,vehicle path planning algorithm is studied based on RGB-D SLAM in unknown environment,and a set of path planning algorithm based on reinforcement learning is constructed.This paper mainly focuses on the following three parts:1.The Design and Implementation of RGB-D SLAM System.This paper introduces image feature extraction and matching,robot motion estimation,optimization and mapping,and constructs a routine RGB-D SLAM system,through which the three-dimensional point cloud map of the environment can be constructed.2.Octomap Modeling and Map Conversion.In this part,octree maps are introduced in detail,and by using the three-dimensional point cloud information output from SLAM system as input,Octomap modeling and three-dimensional projection transformation are realized,through which Point cloud maps with large scale and large space can be transformed into two-dimensional raster maps successfully.3.The research on path planning based on reinforcement learning.This part is the focus of the full paper.The SARSA algorithm,Q-Learning algorithm and Deep Q-Learning algorithm are tested and compared in order to find the most suitable algorithm for this paper.Through the construction of Depp Q-Learning algorithm,the feasibility of the algorithm is confirmed by the test run in the simulation environment.Finally,the algorithm is run in a real robot to verify the effectiveness of the algorithm.Above all,this paper uses RGB-D SLAM to reconstruct three-dimensional point clouds as environmental input,and uses Deep Q-Learning algorithm to realize autonomous location and path planning of mobile robots.